A Deep Learning-based in silico Framework for Optimization on Retinal Prosthetic Stimulation.

Area of Science:

• Biomedical Engineering
• Computational Neuroscience
• Artificial Intelligence

Background:

• Retinal implants aim to restore vision by stimulating remaining retinal neurons.
• Accurate simulation of visual perception is crucial for designing effective retinal prosthetics.
• Existing models like pulse2percept offer biomimetic simulations but lack optimization for specific electrode configurations.

Purpose of the Study:

• To develop a neural network-based framework for optimizing visual perceptions simulated by the pulse2percept model.
• To enhance the quality of visual stimuli generated for retinal implant users.
• To enable end-to-end fine-tuning of perception quality through gradient descent.

Main Methods:

• A U-Net convolutional neural network served as a trainable encoder, converting original images into stimuli.
• A pre-trained U-Net mimicked the pulse2percept biomimetic model.
• A VGG classifier evaluated perception quality using original images.
• The framework was tested on 10,000 MNIST dataset images with a 6x10 electrode configuration.

Main Results:

• The neural network encoder significantly outperformed a trivial downsampling approach.
• A 36.17% boost in weighted F1-Score was achieved in the pre-trained classifier.
• The framework demonstrated effective optimization of simulated visual perception.

Conclusions:

• A fully neural network-based encoder can significantly improve simulated visual perception for retinal implants.
• End-to-end optimization using gradient descent allows for fine-tuning perception quality.
• This framework offers a promising approach for developing more effective visual prosthetics.